Shadow-Frog: Coding Agents that Dream and Discover

Microsoft's Shadow-Frog is an agentic discovery system that enhances coding agents by building a codebase memory through active exploration. Instead of passively recording past tasks, Shadow-Frog agents conduct experiments on under-explored parts of a codebase during idle times, distilling their learnings into a structured shadow knowledge base. This `.shadow/` directory mirrors the source tree, allowing agents to proactively find bugs, anticipate features, and improve knowledge retrieval, significantly outperforming baselines in various tasks.